Last quarter Silvertip ran two campaigns for the same client. Same product, same offer, the same creative budget split down the middle. One was Advantage+ with broad targeting and nine creatives. The other was the structure we’d been running for three years: eight ad sets, layered interests, careful exclusions. The broad campaign returned 41% more revenue.
Two years ago we’d have called that a fluke and re-tested it. We’d have been wrong. The auction underneath those campaigns is no longer the auction we learned to optimise against. The name Meta gave its new ranking system is Andromeda, and most of the briefing notes around it are vague. The real story is more interesting than the press release.
What Andromeda actually is, in plain English
Andromeda is Meta’s new deep-learning ad retrieval and ranking system. It sits between every advertiser’s pool of ads and every user’s feed, scoring millions of ad-impression combinations against tens of thousands of features in milliseconds, using Meta’s custom MTIA chips. Where the old system shortlisted a few hundred candidate ads per impression, Andromeda screens the entire eligible inventory. The shortlist is the auction.
That’s the 50-word version. The longer one: Meta has rebuilt the part of the ad system that decides which ad to even consider showing you. Once that funnel-top is wider and smarter, the rest of the system (bidding, pacing, optimisation) gets to work with a much better starting hand.
How it differs from previous Meta optimisation
The pre-Andromeda system was a relevance-and-bid contest with a shortlist. Meta picked a few hundred candidate ads, then ran an auction. Targeting parameters, lookalikes, interest stacks. They were really just ways of saying “make sure the shortlist contains ads relevant to this kind of person.”
Andromeda doesn’t need much of a shortlist. It evaluates the whole inventory of eligible ads in your account (and everyone else’s competing for that impression) at a depth the old system couldn’t.
That changes what targeting is. Audience signals used to be filters. Now they’re inputs, useful for warming up a cold campaign, but increasingly redundant once Meta has seen enough conversions to model user-ad fit directly. This is why the “broad beats granular” finding keeps showing up in case studies that would have been heretical in 2022.
What levers you lose, what you gain
You lose most of the control levers I built my career on. Tight interest stacks. Ad-set–level bid adjustments. The ability to “force” delivery toward an audience by structuring around it. Andromeda quietly overrules that work. If your ad performs better against a slightly different cohort, that’s where it goes, regardless of how clever your audience build looked in the briefing deck.
You gain a system that finds pockets of demand you’d never have built audiences around. I’ve watched Advantage+ campaigns spend 60% of their budget on user clusters no targeting setup I’d have written would have selected. Those users converted at a 22% better CPA than the audiences I’d hand-built next to them. That’s not an edge case. That’s the system working.
The trade is real. Control for performance. If you don’t trust performance to justify lost control, you’ll fight the system. And you’ll lose.
The honest answer on CPA, ROAS, and efficiency at scale
This is the part the press release glosses over. Andromeda doesn’t reliably improve every metric for every account. What it does, consistently, across the accounts I’ve watched migrate over, is widen the gap between accounts with strong inputs and accounts without them.
If your conversion signal is clean, your creative is varied, and your offer holds up at scale, Andromeda will compound those advantages. Across my own portfolio I’ve seen CPA improvements of 10–25% in the first 60 days post-restructure, with the bigger gains skewing to accounts spending north of £75k a month, where the model has more data to work with.
If your signal is patchy or your creative pool is thin, the same system will surface those weaknesses faster than the old one did. I’ve watched two accounts get measurably worse under Andromeda, not because the model is broken, but because it stripped away the structural hiding spots that were masking weak inputs.
So the real question isn’t “does Andromeda improve ROAS?” The real question is whether your inputs are good enough for a smarter system to find the upside in them.
What this changes about how you operate
Stop optimising the campaign. Start optimising the inputs. The campaign is doing what Meta tells it to. You don’t have to like it.
In practice that means fewer ad sets, more creative variants, cleaner conversion tracking, and a higher bar for what you let into the account in the first place. The agencies still building 14-ad-set campaigns with five layered audiences are running yesterday’s playbook on this year’s system. They’ll show you healthy CTRs and ad-set–level ROAS that look fine in isolation, and a blended business that’s quietly underperforming because the structure is fighting the model.
If your last meaningful Meta restructure was before mid-2024, you’re almost certainly leaving money on the table. That’s the simplest read on Andromeda. The rest is detail.
If you’re trying to work out whether your Meta setup is keeping up with what’s underneath it, I’m easy to reach.
